Re: [PATCH v2 5/5] arm64: perf: Add support for chaining event counters

[Mark Rutland <mark.rutland@arm.com>]

On Mon, Jun 11, 2018 at 02:54:16PM +0100, Suzuki K Poulose wrote:
> On 08/06/18 15:46, Suzuki K Poulose wrote:
> > On 06/06/2018 07:01 PM, Mark Rutland wrote:
> > > On Tue, May 29, 2018 at 11:55:56AM +0100, Suzuki K Poulose wrote:
> 
> > > > -        value |= 0xffffffff00000000ULL;
> > > > +        if (!armv8pmu_event_is_64bit(event))
> > > > +            value |= 0xffffffff00000000ULL;
> > > >           write_sysreg(value, pmccntr_el0);
> > > > -    } else if (armv8pmu_select_counter(idx) == idx)
> > > > -        write_sysreg(value, pmxevcntr_el0);
> > > > +    } else
> > > > +        armv8pmu_write_hw_counter(event, value);
> > > >   }
> > > 
> > > > +static inline void armv8pmu_write_event_type(struct perf_event *event)
> > > > +{
> > > > +    struct hw_perf_event *hwc = &event->hw;
> > > > +    int idx = hwc->idx;
> > > > +
> > > > +    /*
> > > > +     * For chained events, write the the low counter event type
> > > > +     * followed by the high counter. The high counter is programmed
> > > > +     * with CHAIN event code with filters set to count at all ELs.
> > > > +     */
> > > > +    if (armv8pmu_event_is_chained(event)) {
> > > > +        u32 chain_evt = ARMV8_PMUV3_PERFCTR_CHAIN |
> > > > +                ARMV8_PMU_INCLUDE_EL2;
> > > > +
> > > > +        armv8pmu_write_evtype(idx - 1, hwc->config_base);
> > > > +        isb();
> > > > +        armv8pmu_write_evtype(idx, chain_evt);
> > > 
> > > The ISB isn't necessary here, AFAICT. We only do this while the PMU is
> > > disabled; no?
> > 
> > You're right. I was just following the ARM ARM.
> 
> Taking another look, it is not clear about the semantics of "pmu->enable()"
> and pmu->disable() callbacks. 

I was talking about pmu::{pmu_disable,pmu_enable}(), so I'm not sure I
follow how arm_pmu::{enable,disable}() are relevant here.

The arm_pmu::{enable,disable}() callbacks enable or disable individual
counters. For example, leaving unused counters disabled may save power,
even if the PMU as a whole is enabled.

> I don't see any reference to them in the perf core
> driver anymore. The perf core uses add() / del () instead, with the PMU
> turned off. Do you have any idea about the enable()/disable() callbacks ?

I'm not sure I understand what you're asking here.

Thanks,
Mark.